tech
China's latest open-weight model rivals U.S. models at hacking
The model is expected to be released widely in two weeks
TL;DR
- Chinese AI lab Z.ai's GLM-5.3 model demonstrates strong capabilities in finding and exploiting security flaws, rivaling U.S. frontier models.
- GLM-5.3 scored 84.5% on CyberGym, a benchmark for identifying security vulnerabilities, outperforming models from Anthropic and OpenAI.
- Z.ai is delaying the public release of GLM-5.3 for two weeks to implement and test enhanced safety and security controls.
- The company is implementing a tiered access program for selected security partners to use GLM-5.3 in controlled environments.
- Z.ai frames GLM-5.3 as a 'shield' for defenders, aiming to help secure open-source projects, and has launched a program for open-source maintainers to scan their repositories for bugs.
- Concerns exist regarding the potential misuse of GLM-5.3 once its weights are publicly released, as Z.ai will not be able to control modifications or usage.
- Recent incidents highlight the growing use of open-source AI agents in cyberattacks, such as an automated attack against Taiwan's government.